CoreWeave looks to Wall Street to hedge memory chips, an asset with no market

Memory chips have roughly doubled in price this year, a squeeze that has emptied Apple’s shelves and rewritten data centre budgets. CoreWeave is now working out what to do if they get cheaper.

The AI cloud company is exploring financial derivatives as a hedge against a future drop in memory and storage chip prices, Reuters reported on July 14, citing a person familiar with the matter. The talks are early, no hedges have been executed, and no CoreWeave executive has discussed them publicly. The account rests on a single unnamed source.

The exposure is a by-product of guarding against the opposite problem. Cloud operators including CoreWeave have locked in supply through long-term agreements with memory and storage makers such as Micron and SanDisk, many of which guarantee the supplier a price floor on DRAM and storage chips.

That shields the chipmaker from a downturn and leaves the buyer paying well above the going rate if prices fall. Memory has always been cyclical, and elevated prices have a habit of collapsing once new capacity switches on.

The numbers behind the anxiety belong to TrendForce. Conventional DRAM contract prices rose 93% to 98% quarter on quarter in the first quarter, lifting industry revenue 81% to $97 billion, and were forecast to climb another 58% to 63% in the second.

The 💜 of EU tech

The latest rumblings from the EU tech scene, a story from our wise ol' founder Boris, and some questionable AI art. It's free, every week, in your inbox. Sign up now!

The climb is decelerating, and the long-term agreements are the reason. TrendForce expects server DRAM contract prices to rise 13% to 18% in the third quarter, with most of the remaining increase landing on buyers who never signed one.

Those agreements are collars in all but name: a floor for the seller, a ceiling for the buyer. SK hynix has reportedly stripped the cap out of its contracts, the only major supplier said to have done so, while terms stretch from one year to three or five.

CoreWeave’s stake in the outcome is a 2026 capital expenditure budget of $31bn to $35bn, whose low end it raised in May because of component prices, finance chief Nitin Agrawal told analysts. The buildout is financed largely with debt, though borrowing costs have fallen from 10% to 7% in six months.

“It’s an issue, it’s a problem, but we have an incredible capacity to navigate the supply chain,” chief executive Mike Intrator said on the same call. Demand is not the worry. CoreWeave closed the quarter with a $99.4bn backlog and has since added customers including Jane Street, on a $6bn cloud deal.

What CoreWeave would actually trade is the harder question. Exchanges have spent the year building a market for compute, but the CME Group and Intercontinental Exchange contracts track GPU rental rates rather than memory, and the first futures to reference DRAM, announced by Architect Financial Technologies and index provider Ornn in January, are still waiting on regulatory approval.

Which leaves equities. The instruments under discussion reportedly include put options, and Reuters described the risk being hedged as a slide in memory chip stocks, wording that suggests the trade is a proxy rather than a bet on the chips themselves. The liquid way to be short DRAM today is to be short the people who make it.

There is a ready-made vehicle for that. The Roundhill Memory ETF, ticker DRAM, launched on April 2 and pulled in $6.5bn within 27 trading days, with roughly three-quarters of the fund in SK hynix, Samsung, and Micron. Micron and SanDisk, the two suppliers named in CoreWeave’s long-term deals, are both holdings.

Hedging an input is an old idea, and not a reliably good one. Airlines have been burned by fuel hedges before, a comparison the Reuters account makes itself.

The cycle CoreWeave is bracing for has a date attached. SK hynix and Micron have both signalled that their new manufacturing capacity will be fully ramped in early 2028, which is when the price of a chip CoreWeave has already agreed to buy stops being a hypothetical.